Lightweight Object Detection Network With Spatial-Channel Attention and Stair Fusion Structure for Remote Sensing Open-Pit

Effective monitoring methods for open-pit mining areas are important safeguards for mineral resource and environmental protection. Due to the complex surface environment of mining areas and the difficulty in defining the edges, conventional methods are difficult to detect accurately. Furthermore, previous deep-learning-based approaches have high memory and computational costs. Therefore, we propose a novel lightweight network with dual-attention and stair fusion structure called DASFNet for pixel-level object detection in remote sensing mining areas. First, we design a neighbor feature aggregation module (NFAM) to integrate adjacent features, which can enhance the representation of features at each level. Second, an edge feature compensation module (EFCM) is developed to strengthen edge information and improve the detection performance of mining area edges. In addition, a cross-attention fusion module (CAFM) is developed to weight features from adjacent levels, with the aim of better aggregating features from high to low levels. Finally, we create and introduce a dataset called CIGMR for open-pit mining area detection. To our knowledge, CIGMR is the first dataset for open-pit mining area detection. At the same time, by applying fog simulation algorithms to the CIGMR dataset, we obtained the CIGMR(fog) dataset. Extensive experiments demonstrate that DASFNet is superior to six state-of-the-art (SOTA) methods on the CIGMR and CIGMR(fog) datasets, which has fewer parameters and lower computational cost.